About Positron

A visionary, an applied mathematician, and an engineer walk into a bar
(our founding story)

We started Positron for one reason: to stop GPUs from bankrupting everyone trying to deploy AI.

Turns out, inference—the unsexy side of AI that powers billions of daily model queries—is now devouring budgets, frying electrical grids, and giving CTOs (and CFOs) migraines.

Our mission is simple and urgent: deliver inference that's ridiculously efficient, surprisingly affordable, proudly made in America—and (finally!) makes GPUs optional.

The founders of Positron, and everyone who has joined since, can see that GPUs for inference are, well, CRAP*, and that there is a better, more efficient way to run the LLMs and AIs that we are all increasingly depending on. We see an exciting and significant opportunity before us to address this.

* our lawyers would like to point out that CRAP simply means: Chips Reaping Absurd Profits and has nothing to do with subjective quality or barnyard slang.

Everyone at Positron has their own unique intuition for this, and so over many nights, in many bars and hotels across the western U.S., through many napkin diagrams, countless disagreements (no fist fights; one pen thrown), and a few really key executable technical insights, Positron’s first product, Atlas, was created on one of those napkins. We have since lost the napkin, but you can buy Atlas today.

Atlas is the world’s first LLM-inference-first accelerator, built in just 18 months from conception, manufactured, and fully supported right here in the U.S., designed with efficiency in mind at a time when the entire AI sector (which, let's face it, is most of the tech industry right now) is grappling with cost and energy constraints that are stalling growth. Unless, that is, you're lucky enough to be one of a few trillion-dollar companies who knows how to reboot an old nuclear power plant.

But this is just the introduction to our story. We're building an even faster, more efficient second-generation system called Titan — four to eight Asimov chips and up to 18.4 TB of directly accelerator-attached memory, built for the agentic, long-context, and multimodal workloads that current GPU infrastructure can't sustain. The 'napkin' for this one is Atlas itself (it's harder to lose deployed Atlas servers). We're taking every piece of learning and insight Atlas gives us to deliver an inference solution that unlocks near-limitless context and the ability to concurrently hold large numbers of models and AI agents ready to instantly query — all enabled by terabytes of per-accelerator memory. And we'll deliver it in a form factor that doesn't require expensive liquid cooling, ridiculously over-provisioned boutique networking, or force our customers to upgrade their infrastructure.

That last bit does sound like bar talk - a too-good-to-be-true elevator pitch. To ground this, we aren't claiming that we invented a fundamental new technology: “let me tell you all about in-memory-analog-wafer-sized-optical-quantum-mram-deterministic-approximate-inference”. No, that’s not us. But we are claiming that we have launched Positron on a trajectory that has earned us the right to crack open the inference market. By rapidly iterating on Atlas, we are taking a systematic approach to reassess and redefine what it takes to run LLMs and transformers. We are building inference focused silicon and a technology platform, informed by customers, that creates allies and partners, enabling all of us to break the vertical stranglehold that a certain pair of closely related GPUs force upon the market to overcome their fundamental inefficiencies.

As we have often explained to many bartenders, “we are just trying to save businesses from drowning in GPU invoices”. And they get it because the mocktail they just served was from a recipe made by ChatGPT, and their cousin in Florida told them to invest in Nvidia.